1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
6 * This file is part of Lustre, http://www.lustre.org.
8 * Lustre is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Lustre is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Lustre; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 #define DEBUG_SUBSYSTEM S_RPC
27 #include <liblustre.h>
30 #include <linux/obd_support.h>
31 #include <linux/obd_class.h>
32 #include <linux/lustre_lib.h>
33 #include <linux/lustre_ha.h>
34 #include <linux/lustre_import.h>
36 #include "ptlrpc_internal.h"
38 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
39 struct ptlrpc_client *cl)
41 cl->cli_request_portal = req_portal;
42 cl->cli_reply_portal = rep_portal;
46 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
48 struct ptlrpc_connection *c;
49 struct ptlrpc_peer peer;
52 err = ptlrpc_uuid_to_peer(uuid, &peer);
54 CERROR("cannot find peer %s!\n", uuid->uuid);
58 c = ptlrpc_get_connection(&peer, uuid);
60 memcpy(c->c_remote_uuid.uuid,
61 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
64 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
69 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
70 struct obd_uuid *uuid)
72 struct ptlrpc_peer peer;
75 err = ptlrpc_uuid_to_peer(uuid, &peer);
77 CERROR("cannot find peer %s!\n", uuid->uuid);
81 memcpy(&conn->c_peer, &peer, sizeof (peer));
85 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
87 struct ptlrpc_bulk_desc *desc;
89 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
93 spin_lock_init(&desc->bd_lock);
94 init_waitqueue_head(&desc->bd_waitq);
95 desc->bd_max_iov = npages;
96 desc->bd_iov_count = 0;
97 desc->bd_md_h = PTL_INVALID_HANDLE;
98 desc->bd_portal = portal;
104 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp (struct ptlrpc_request *req,
105 int npages, int type, int portal)
107 struct obd_import *imp = req->rq_import;
108 struct ptlrpc_bulk_desc *desc;
110 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
111 desc = new_bulk(npages, type, portal);
115 desc->bd_import_generation = req->rq_import_generation;
116 desc->bd_import = class_import_get(imp);
119 desc->bd_cbid.cbid_fn = client_bulk_callback;
120 desc->bd_cbid.cbid_arg = desc;
122 /* This makes req own desc, and free it when she frees herself */
128 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp (struct ptlrpc_request *req,
129 int npages, int type, int portal)
131 struct obd_export *exp = req->rq_export;
132 struct ptlrpc_bulk_desc *desc;
134 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
136 desc = new_bulk(npages, type, portal);
140 desc->bd_export = class_export_get(exp);
143 desc->bd_cbid.cbid_fn = server_bulk_callback;
144 desc->bd_cbid.cbid_arg = desc;
146 /* NB we don't assign rq_bulk here; server-side requests are
147 * re-used, and the handler frees the bulk desc explicitly. */
152 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
153 struct page *page, int pageoffset, int len)
155 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
156 LASSERT(page != NULL);
157 LASSERT(pageoffset >= 0);
159 LASSERT(pageoffset + len <= PAGE_SIZE);
163 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
166 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
170 LASSERT(desc != NULL);
171 LASSERT(desc->bd_iov_count != 0x5a5a5a5a); /* not freed already */
172 LASSERT(!desc->bd_network_rw); /* network hands off or */
173 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
175 class_export_put(desc->bd_export);
177 class_import_put(desc->bd_import);
179 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
180 bd_iov[desc->bd_max_iov]));
184 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, int opcode,
185 int count, int *lengths, char **bufs)
187 struct ptlrpc_request *request;
191 LASSERT((unsigned long)imp > 0x1000);
193 OBD_ALLOC(request, sizeof(*request));
195 CERROR("request allocation out of memory\n");
199 rc = lustre_pack_request(request, count, lengths, bufs);
201 CERROR("cannot pack request %d\n", rc);
202 OBD_FREE(request, sizeof(*request));
206 if (imp->imp_server_timeout)
207 request->rq_timeout = obd_timeout / 2;
209 request->rq_timeout = obd_timeout;
210 request->rq_send_state = LUSTRE_IMP_FULL;
211 request->rq_type = PTL_RPC_MSG_REQUEST;
212 request->rq_import = class_import_get(imp);
214 request->rq_req_cbid.cbid_fn = request_out_callback;
215 request->rq_req_cbid.cbid_arg = request;
217 request->rq_reply_cbid.cbid_fn = reply_in_callback;
218 request->rq_reply_cbid.cbid_arg = request;
220 request->rq_phase = RQ_PHASE_NEW;
222 /* XXX FIXME bug 249 */
223 request->rq_request_portal = imp->imp_client->cli_request_portal;
224 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
226 spin_lock_init(&request->rq_lock);
227 INIT_LIST_HEAD(&request->rq_list);
228 INIT_LIST_HEAD(&request->rq_replay_list);
229 INIT_LIST_HEAD(&request->rq_set_chain);
230 init_waitqueue_head(&request->rq_reply_waitq);
231 request->rq_xid = ptlrpc_next_xid();
232 atomic_set(&request->rq_refcount, 1);
234 request->rq_reqmsg->opc = opcode;
235 request->rq_reqmsg->flags = 0;
240 struct ptlrpc_request_set *ptlrpc_prep_set(void)
242 struct ptlrpc_request_set *set;
244 OBD_ALLOC(set, sizeof *set);
247 INIT_LIST_HEAD(&set->set_requests);
248 init_waitqueue_head(&set->set_waitq);
249 set->set_remaining = 0;
250 spin_lock_init(&set->set_new_req_lock);
251 INIT_LIST_HEAD(&set->set_new_requests);
256 /* Finish with this set; opposite of prep_set. */
257 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
259 struct list_head *tmp;
260 struct list_head *next;
265 /* Requests on the set should either all be completed, or all be new */
266 expected_phase = (set->set_remaining == 0) ?
267 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
268 list_for_each (tmp, &set->set_requests) {
269 struct ptlrpc_request *req =
270 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
272 LASSERT(req->rq_phase == expected_phase);
276 LASSERT(set->set_remaining == 0 || set->set_remaining == n);
278 list_for_each_safe(tmp, next, &set->set_requests) {
279 struct ptlrpc_request *req =
280 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
281 list_del_init(&req->rq_set_chain);
283 LASSERT(req->rq_phase == expected_phase);
285 if (req->rq_phase == RQ_PHASE_NEW) {
287 if (req->rq_interpret_reply != NULL) {
288 int (*interpreter)(struct ptlrpc_request *,
290 req->rq_interpret_reply;
292 /* higher level (i.e. LOV) failed;
293 * let the sub reqs clean up */
294 req->rq_status = -EBADR;
295 interpreter(req, &req->rq_async_args,
298 set->set_remaining--;
302 ptlrpc_req_finished (req);
305 LASSERT(set->set_remaining == 0);
307 OBD_FREE(set, sizeof(*set));
311 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
312 struct ptlrpc_request *req)
314 /* The set takes over the caller's request reference */
315 list_add_tail(&req->rq_set_chain, &set->set_requests);
317 set->set_remaining++;
318 atomic_inc(&req->rq_import->imp_inflight);
321 /* lock so many callers can add things, the context that owns the set
322 * is supposed to notice these and move them into the set proper. */
323 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
324 struct ptlrpc_request *req)
327 spin_lock_irqsave(&set->set_new_req_lock, flags);
328 /* The set takes over the caller's request reference */
329 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
331 spin_unlock_irqrestore(&set->set_new_req_lock, flags);
335 * Based on the current state of the import, determine if the request
336 * can be sent, is an error, or should be delayed.
338 * Returns true if this request should be delayed. If false, and
339 * *status is set, then the request can not be sent and *status is the
340 * error code. If false and status is 0, then request can be sent.
342 * The imp->imp_lock must be held.
344 static int ptlrpc_import_delay_req(struct obd_import *imp,
345 struct ptlrpc_request *req, int *status)
350 LASSERT (status != NULL);
353 if (imp->imp_state == LUSTRE_IMP_NEW) {
354 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
358 else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
359 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
362 /* allow CONNECT even if import is invalid */
363 else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
364 imp->imp_state == LUSTRE_IMP_CONNECTING) {
368 * If the import has been invalidated (such as by an OST failure), the
369 * request must fail with -EIO.
371 else if (imp->imp_invalid) {
372 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
375 else if (req->rq_import_generation != imp->imp_generation) {
376 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
379 else if (req->rq_send_state != imp->imp_state) {
380 if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake
382 *status = -EWOULDBLOCK;
390 static int ptlrpc_check_reply(struct ptlrpc_request *req)
396 /* serialise with network callback */
397 spin_lock_irqsave (&req->rq_lock, flags);
399 if (req->rq_replied) {
400 DEBUG_REQ(D_NET, req, "REPLIED:");
405 DEBUG_REQ(D_ERROR, req, "ABORTED:");
409 if (req->rq_resend) {
410 DEBUG_REQ(D_ERROR, req, "RESEND:");
414 if (req->rq_restart) {
415 DEBUG_REQ(D_ERROR, req, "RESTART:");
420 spin_unlock_irqrestore (&req->rq_lock, flags);
421 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
425 static int ptlrpc_check_status(struct ptlrpc_request *req)
430 err = req->rq_repmsg->status;
431 if (req->rq_repmsg->type == PTL_RPC_MSG_ERR) {
432 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err == %d",
434 RETURN(err < 0 ? err : -EINVAL);
438 DEBUG_REQ(D_INFO, req, "status is %d", err);
439 } else if (err > 0) {
440 /* XXX: translate this error from net to host */
441 DEBUG_REQ(D_INFO, req, "status is %d", err);
447 static int after_reply(struct ptlrpc_request *req)
450 struct obd_import *imp = req->rq_import;
454 LASSERT(!req->rq_receiving_reply);
456 /* NB Until this point, the whole of the incoming message,
457 * including buflens, status etc is in the sender's byte order. */
460 /* Clear reply swab mask; this is a new reply in sender's byte order */
461 req->rq_rep_swab_mask = 0;
463 LASSERT (req->rq_nob_received <= req->rq_replen);
464 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
466 CERROR("unpack_rep failed: %d\n", rc);
470 if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
471 req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
472 CERROR("invalid packet type received (type=%u)\n",
473 req->rq_repmsg->type);
477 /* Store transno in reqmsg for replay. */
478 req->rq_reqmsg->transno = req->rq_transno = req->rq_repmsg->transno;
480 rc = ptlrpc_check_status(req);
482 /* Either we've been evicted, or the server has failed for
483 * some reason. Try to reconnect, and if that fails, punt to the
485 if (rc == -ENOTCONN) {
486 if (req->rq_send_state != LUSTRE_IMP_FULL ||
487 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
491 ptlrpc_request_handle_notconn(req);
496 if (req->rq_import->imp_replayable) {
497 spin_lock_irqsave(&imp->imp_lock, flags);
498 if (req->rq_replay || req->rq_transno != 0)
499 ptlrpc_retain_replayable_request(req, imp);
500 else if (req->rq_commit_cb != NULL) {
501 spin_unlock_irqrestore(&imp->imp_lock, flags);
502 req->rq_commit_cb(req);
503 spin_lock_irqsave(&imp->imp_lock, flags);
506 if (req->rq_transno > imp->imp_max_transno)
507 imp->imp_max_transno = req->rq_transno;
509 /* Replay-enabled imports return commit-status information. */
510 if (req->rq_repmsg->last_committed)
511 imp->imp_peer_committed_transno =
512 req->rq_repmsg->last_committed;
513 ptlrpc_free_committed(imp);
514 spin_unlock_irqrestore(&imp->imp_lock, flags);
520 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
522 struct obd_import *imp;
527 LASSERT(req->rq_phase == RQ_PHASE_NEW);
528 req->rq_phase = RQ_PHASE_RPC;
530 imp = req->rq_import;
531 spin_lock_irqsave(&imp->imp_lock, flags);
533 req->rq_import_generation = imp->imp_generation;
535 if (ptlrpc_import_delay_req(imp, req, &rc)) {
536 spin_lock (&req->rq_lock);
538 spin_unlock (&req->rq_lock);
540 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
542 req->rq_reqmsg->status,
543 ptlrpc_import_state_name(req->rq_send_state),
544 ptlrpc_import_state_name(imp->imp_state));
545 LASSERT(list_empty (&req->rq_list));
547 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
548 spin_unlock_irqrestore(&imp->imp_lock, flags);
553 spin_unlock_irqrestore(&imp->imp_lock, flags);
555 req->rq_phase = RQ_PHASE_INTERPRET;
559 /* XXX this is the same as ptlrpc_queue_wait */
560 LASSERT(list_empty(&req->rq_list));
561 list_add_tail(&req->rq_list, &imp->imp_sending_list);
562 spin_unlock_irqrestore(&imp->imp_lock, flags);
564 req->rq_reqmsg->status = current->pid;
565 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc"
566 " %s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
567 imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
569 imp->imp_connection->c_peer.peer_ni->pni_name,
570 imp->imp_connection->c_peer.peer_nid,
571 req->rq_reqmsg->opc);
573 rc = ptl_send_rpc(req);
575 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
582 int ptlrpc_check_set(struct ptlrpc_request_set *set)
585 struct list_head *tmp;
586 int force_timer_recalc = 0;
589 if (set->set_remaining == 0)
592 list_for_each(tmp, &set->set_requests) {
593 struct ptlrpc_request *req =
594 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
595 struct obd_import *imp = req->rq_import;
598 if (req->rq_phase == RQ_PHASE_NEW &&
599 ptlrpc_send_new_req(req)) {
600 force_timer_recalc = 1;
603 if (!(req->rq_phase == RQ_PHASE_RPC ||
604 req->rq_phase == RQ_PHASE_BULK ||
605 req->rq_phase == RQ_PHASE_INTERPRET ||
606 req->rq_phase == RQ_PHASE_COMPLETE)) {
607 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
611 if (req->rq_phase == RQ_PHASE_COMPLETE)
614 if (req->rq_phase == RQ_PHASE_INTERPRET)
615 GOTO(interpret, req->rq_status);
618 ptlrpc_unregister_reply(req);
619 if (req->rq_status == 0)
620 req->rq_status = -EIO;
621 req->rq_phase = RQ_PHASE_INTERPRET;
623 spin_lock_irqsave(&imp->imp_lock, flags);
624 list_del_init(&req->rq_list);
625 spin_unlock_irqrestore(&imp->imp_lock, flags);
627 GOTO(interpret, req->rq_status);
630 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
631 * will only be set after rq_timedout, but the oig waiting
632 * path sets rq_intr irrespective of whether ptlrpcd has
633 * seen a timeout. our policy is to only interpret
634 * interrupted rpcs after they have timed out */
635 if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
636 /* NB could be on delayed list */
637 ptlrpc_unregister_reply(req);
638 req->rq_status = -EINTR;
639 req->rq_phase = RQ_PHASE_INTERPRET;
641 spin_lock_irqsave(&imp->imp_lock, flags);
642 list_del_init(&req->rq_list);
643 spin_unlock_irqrestore(&imp->imp_lock, flags);
645 GOTO(interpret, req->rq_status);
648 if (req->rq_phase == RQ_PHASE_RPC) {
649 if (req->rq_waiting || req->rq_resend) {
652 LASSERT (!ptlrpc_client_receiving_reply(req));
653 LASSERT (req->rq_bulk == NULL ||
654 !ptlrpc_bulk_active(req->rq_bulk));
656 spin_lock_irqsave(&imp->imp_lock, flags);
658 if (ptlrpc_import_delay_req(imp, req, &status)) {
659 spin_unlock_irqrestore(&imp->imp_lock,
664 list_del_init(&req->rq_list);
666 req->rq_status = status;
667 req->rq_phase = RQ_PHASE_INTERPRET;
668 spin_unlock_irqrestore(&imp->imp_lock,
670 GOTO(interpret, req->rq_status);
672 if (req->rq_no_resend) {
673 req->rq_status = -ENOTCONN;
674 req->rq_phase = RQ_PHASE_INTERPRET;
675 spin_unlock_irqrestore(&imp->imp_lock,
677 GOTO(interpret, req->rq_status);
679 list_add_tail(&req->rq_list,
680 &imp->imp_sending_list);
682 spin_unlock_irqrestore(&imp->imp_lock, flags);
685 if (req->rq_resend) {
686 lustre_msg_add_flags(req->rq_reqmsg,
688 ptlrpc_unregister_reply(req);
690 __u64 old_xid = req->rq_xid;
692 /* ensure previous bulk fails */
693 req->rq_xid = ptlrpc_next_xid();
694 CDEBUG(D_HA, "resend bulk "
697 old_xid, req->rq_xid);
701 rc = ptl_send_rpc(req);
703 DEBUG_REQ(D_HA, req, "send failed (%d)",
705 force_timer_recalc = 1;
708 /* need to reset the timeout */
709 force_timer_recalc = 1;
712 /* Still waiting for a reply? */
713 if (ptlrpc_client_receiving_reply(req))
716 /* Did we actually receive a reply? */
717 if (!ptlrpc_client_replied(req))
720 spin_lock_irqsave(&imp->imp_lock, flags);
721 list_del_init(&req->rq_list);
722 spin_unlock_irqrestore(&imp->imp_lock, flags);
724 req->rq_status = after_reply(req);
725 if (req->rq_resend) {
726 /* Add this req to the delayed list so
727 it can be errored if the import is
728 evicted after recovery. */
729 spin_lock_irqsave (&req->rq_lock, flags);
730 list_add_tail(&req->rq_list,
731 &imp->imp_delayed_list);
732 spin_unlock_irqrestore(&req->rq_lock, flags);
736 /* If there is no bulk associated with this request,
737 * then we're done and should let the interpreter
738 * process the reply. Similarly if the RPC returned
739 * an error, and therefore the bulk will never arrive.
741 if (req->rq_bulk == NULL || req->rq_status != 0) {
742 req->rq_phase = RQ_PHASE_INTERPRET;
743 GOTO(interpret, req->rq_status);
746 req->rq_phase = RQ_PHASE_BULK;
749 LASSERT(req->rq_phase == RQ_PHASE_BULK);
750 if (ptlrpc_bulk_active(req->rq_bulk))
753 if (!req->rq_bulk->bd_success) {
754 /* The RPC reply arrived OK, but the bulk screwed
755 * up! Dead wierd since the server told us the RPC
756 * was good after getting the REPLY for her GET or
757 * the ACK for her PUT. */
758 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
762 req->rq_phase = RQ_PHASE_INTERPRET;
765 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
766 LASSERT(!req->rq_receiving_reply);
768 ptlrpc_unregister_reply(req);
769 if (req->rq_bulk != NULL)
770 ptlrpc_unregister_bulk (req);
772 req->rq_phase = RQ_PHASE_COMPLETE;
774 if (req->rq_interpret_reply != NULL) {
775 int (*interpreter)(struct ptlrpc_request *,void *,int) =
776 req->rq_interpret_reply;
777 req->rq_status = interpreter(req, &req->rq_async_args,
781 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:"
782 "opc %s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
783 imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
785 imp->imp_connection->c_peer.peer_ni->pni_name,
786 imp->imp_connection->c_peer.peer_nid,
787 req->rq_reqmsg->opc);
789 set->set_remaining--;
791 atomic_dec(&imp->imp_inflight);
792 wake_up(&imp->imp_recovery_waitq);
795 /* If we hit an error, we want to recover promptly. */
796 RETURN(set->set_remaining == 0 || force_timer_recalc);
799 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
802 struct obd_import *imp = req->rq_import;
805 DEBUG_REQ(D_ERROR, req, "timeout");
807 spin_lock_irqsave (&req->rq_lock, flags);
808 req->rq_timedout = 1;
809 spin_unlock_irqrestore (&req->rq_lock, flags);
811 ptlrpc_unregister_reply (req);
813 if (req->rq_bulk != NULL)
814 ptlrpc_unregister_bulk (req);
817 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
821 /* The DLM server doesn't want recovery run on its imports. */
822 if (imp->imp_dlm_fake)
825 /* If this request is for recovery or other primordial tasks,
826 * then error it out here. */
827 if (req->rq_send_state != LUSTRE_IMP_FULL ||
828 imp->imp_obd->obd_no_recov) {
829 spin_lock_irqsave (&req->rq_lock, flags);
830 req->rq_status = -ETIMEDOUT;
832 spin_unlock_irqrestore (&req->rq_lock, flags);
836 ptlrpc_fail_import(imp, req->rq_import_generation);
841 int ptlrpc_expired_set(void *data)
843 struct ptlrpc_request_set *set = data;
844 struct list_head *tmp;
845 time_t now = LTIME_S (CURRENT_TIME);
848 LASSERT(set != NULL);
850 /* A timeout expired; see which reqs it applies to... */
851 list_for_each (tmp, &set->set_requests) {
852 struct ptlrpc_request *req =
853 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
855 /* request in-flight? */
856 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting
857 && !req->rq_resend) ||
858 (req->rq_phase == RQ_PHASE_BULK)))
861 if (req->rq_timedout || /* already dealt with */
862 req->rq_sent + req->rq_timeout > now) /* not expired */
865 /* deal with this guy */
866 ptlrpc_expire_one_request (req);
869 /* When waiting for a whole set, we always to break out of the
870 * sleep so we can recalculate the timeout, or enable interrupts
871 * iff everyone's timed out.
876 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
879 spin_lock_irqsave(&req->rq_lock, flags);
881 spin_unlock_irqrestore(&req->rq_lock, flags);
884 void ptlrpc_interrupted_set(void *data)
886 struct ptlrpc_request_set *set = data;
887 struct list_head *tmp;
889 LASSERT(set != NULL);
890 CERROR("INTERRUPTED SET %p\n", set);
892 list_for_each(tmp, &set->set_requests) {
893 struct ptlrpc_request *req =
894 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
896 if (req->rq_phase != RQ_PHASE_RPC)
899 ptlrpc_mark_interrupted(req);
903 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
905 struct list_head *tmp;
906 time_t now = LTIME_S(CURRENT_TIME);
909 struct ptlrpc_request *req;
912 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
914 list_for_each(tmp, &set->set_requests) {
915 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
917 /* request in-flight? */
918 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
919 (req->rq_phase == RQ_PHASE_BULK)))
922 if (req->rq_timedout) /* already timed out */
925 deadline = req->rq_sent + req->rq_timeout;
926 if (deadline <= now) /* actually expired already */
927 timeout = 1; /* ASAP */
928 else if (timeout == 0 || timeout > deadline - now)
929 timeout = deadline - now;
935 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
937 struct list_head *tmp;
938 struct ptlrpc_request *req;
939 struct l_wait_info lwi;
943 LASSERT(!list_empty(&set->set_requests));
944 list_for_each(tmp, &set->set_requests) {
945 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
946 if (req->rq_phase == RQ_PHASE_NEW)
947 (void)ptlrpc_send_new_req(req);
951 timeout = ptlrpc_set_next_timeout(set);
953 /* wait until all complete, interrupted, or an in-flight
955 CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
957 lwi = LWI_TIMEOUT_INTR((timeout ? timeout : 1) * HZ,
959 ptlrpc_interrupted_set, set);
960 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
962 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
964 /* -EINTR => all requests have been flagged rq_intr so next
966 * -ETIMEOUTD => someone timed out. When all reqs have
967 * timed out, signals are enabled allowing completion with
969 * I don't really care if we go once more round the loop in
970 * the error cases -eeb. */
973 LASSERT(set->set_remaining == 0);
976 list_for_each(tmp, &set->set_requests) {
977 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
979 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
980 if (req->rq_status != 0)
984 if (set->set_interpret != NULL) {
985 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
987 rc = interpreter (set, &set->set_args, rc);
993 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
996 if (request == NULL) {
1001 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1002 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1003 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1004 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1006 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1007 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1008 if (request->rq_import != NULL) {
1009 unsigned long flags = 0;
1011 spin_lock_irqsave(&request->rq_import->imp_lock, flags);
1012 list_del_init(&request->rq_replay_list);
1014 spin_unlock_irqrestore(&request->rq_import->imp_lock,
1017 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1019 if (atomic_read(&request->rq_refcount) != 0) {
1020 DEBUG_REQ(D_ERROR, request,
1021 "freeing request with nonzero refcount");
1025 if (request->rq_repmsg != NULL) {
1026 OBD_FREE(request->rq_repmsg, request->rq_replen);
1027 request->rq_repmsg = NULL;
1029 if (request->rq_reqmsg != NULL) {
1030 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
1031 request->rq_reqmsg = NULL;
1033 if (request->rq_export != NULL) {
1034 class_export_put(request->rq_export);
1035 request->rq_export = NULL;
1037 if (request->rq_import != NULL) {
1038 class_import_put(request->rq_import);
1039 request->rq_import = NULL;
1041 if (request->rq_bulk != NULL)
1042 ptlrpc_free_bulk(request->rq_bulk);
1044 OBD_FREE(request, sizeof(*request));
1048 void ptlrpc_free_req(struct ptlrpc_request *request)
1050 __ptlrpc_free_req(request, 0);
1053 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1054 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1056 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1057 (void)__ptlrpc_req_finished(request, 1);
1060 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1063 if (request == NULL)
1066 if (request == LP_POISON ||
1067 request->rq_reqmsg == LP_POISON) {
1068 CERROR("dereferencing freed request (bug 575)\n");
1073 DEBUG_REQ(D_INFO, request, "refcount now %u",
1074 atomic_read(&request->rq_refcount) - 1);
1076 if (atomic_dec_and_test(&request->rq_refcount)) {
1077 __ptlrpc_free_req(request, locked);
1084 void ptlrpc_req_finished(struct ptlrpc_request *request)
1086 __ptlrpc_req_finished(request, 0);
1089 /* Disengage the client's reply buffer from the network
1090 * NB does _NOT_ unregister any client-side bulk.
1091 * IDEMPOTENT, but _not_ safe against concurrent callers.
1092 * The request owner (i.e. the thread doing the I/O) must call...
1094 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1097 wait_queue_head_t *wq;
1098 struct l_wait_info lwi;
1100 LASSERT(!in_interrupt ()); /* might sleep */
1102 if (!ptlrpc_client_receiving_reply(request))
1105 PtlMDUnlink (request->rq_reply_md_h);
1107 /* We have to l_wait_event() whatever the result, to give liblustre
1108 * a chance to run reply_in_callback() */
1110 if (request->rq_set == NULL)
1111 wq = &request->rq_set->set_waitq;
1113 wq = &request->rq_reply_waitq;
1116 /* Network access will complete in finite time but the HUGE
1117 * timeout lets us CWARN for visibility of sluggish NALs */
1118 lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
1119 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1123 LASSERT (rc == -ETIMEDOUT);
1124 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1128 /* caller must hold imp->imp_lock */
1129 void ptlrpc_free_committed(struct obd_import *imp)
1131 struct list_head *tmp, *saved;
1132 struct ptlrpc_request *req;
1133 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1136 LASSERT(imp != NULL);
1138 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1140 CDEBUG(D_HA, "%s: committing for last_committed "LPU64"\n",
1141 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1143 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1144 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1146 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1147 LASSERT(req != last_req);
1150 if (req->rq_import_generation < imp->imp_generation) {
1151 DEBUG_REQ(D_HA, req, "freeing request with old gen");
1155 if (req->rq_replay) {
1156 DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
1160 /* not yet committed */
1161 if (req->rq_transno > imp->imp_peer_committed_transno) {
1162 DEBUG_REQ(D_HA, req, "stopping search");
1166 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
1167 imp->imp_peer_committed_transno);
1169 if (req->rq_commit_cb != NULL)
1170 req->rq_commit_cb(req);
1171 list_del_init(&req->rq_replay_list);
1172 __ptlrpc_req_finished(req, 1);
1179 void ptlrpc_cleanup_client(struct obd_import *imp)
1186 void ptlrpc_resend_req(struct ptlrpc_request *req)
1188 unsigned long flags;
1190 DEBUG_REQ(D_HA, req, "going to resend");
1191 req->rq_reqmsg->handle.cookie = 0;
1192 req->rq_status = -EAGAIN;
1194 spin_lock_irqsave (&req->rq_lock, flags);
1196 req->rq_timedout = 0;
1198 __u64 old_xid = req->rq_xid;
1200 /* ensure previous bulk fails */
1201 req->rq_xid = ptlrpc_next_xid();
1202 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1203 old_xid, req->rq_xid);
1205 ptlrpc_wake_client_req(req);
1206 spin_unlock_irqrestore (&req->rq_lock, flags);
1209 /* XXX: this function and rq_status are currently unused */
1210 void ptlrpc_restart_req(struct ptlrpc_request *req)
1212 unsigned long flags;
1214 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1215 req->rq_status = -ERESTARTSYS;
1217 spin_lock_irqsave (&req->rq_lock, flags);
1218 req->rq_restart = 1;
1219 req->rq_timedout = 0;
1220 ptlrpc_wake_client_req(req);
1221 spin_unlock_irqrestore (&req->rq_lock, flags);
1224 static int expired_request(void *data)
1226 struct ptlrpc_request *req = data;
1229 RETURN(ptlrpc_expire_one_request(req));
1232 static void interrupted_request(void *data)
1234 unsigned long flags;
1236 struct ptlrpc_request *req = data;
1237 DEBUG_REQ(D_HA, req, "request interrupted");
1238 spin_lock_irqsave (&req->rq_lock, flags);
1240 spin_unlock_irqrestore (&req->rq_lock, flags);
1243 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1246 atomic_inc(&req->rq_refcount);
1250 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1251 struct obd_import *imp)
1253 struct list_head *tmp;
1255 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1257 /* don't re-add requests that have been replayed */
1258 if (!list_empty(&req->rq_replay_list))
1261 lustre_msg_add_flags(req->rq_reqmsg,
1264 LASSERT(imp->imp_replayable);
1265 /* Balanced in ptlrpc_free_committed, usually. */
1266 ptlrpc_request_addref(req);
1267 list_for_each_prev(tmp, &imp->imp_replay_list) {
1268 struct ptlrpc_request *iter =
1269 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1271 /* We may have duplicate transnos if we create and then
1272 * open a file, or for closes retained if to match creating
1273 * opens, so use req->rq_xid as a secondary key.
1274 * (See bugs 684, 685, and 428.)
1275 * XXX no longer needed, but all opens need transnos!
1277 if (iter->rq_transno > req->rq_transno)
1280 if (iter->rq_transno == req->rq_transno) {
1281 LASSERT(iter->rq_xid != req->rq_xid);
1282 if (iter->rq_xid > req->rq_xid)
1286 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1290 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1293 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1297 struct l_wait_info lwi;
1298 struct obd_import *imp = req->rq_import;
1299 unsigned long flags;
1303 LASSERT(req->rq_set == NULL);
1304 LASSERT(!req->rq_receiving_reply);
1305 atomic_inc(&imp->imp_inflight);
1307 /* for distributed debugging */
1308 req->rq_reqmsg->status = current->pid;
1309 LASSERT(imp->imp_obd != NULL);
1310 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc "
1311 "%s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
1312 imp->imp_obd->obd_uuid.uuid,
1313 req->rq_reqmsg->status, req->rq_xid,
1314 imp->imp_connection->c_peer.peer_ni->pni_name,
1315 imp->imp_connection->c_peer.peer_nid,
1316 req->rq_reqmsg->opc);
1318 /* Mark phase here for a little debug help */
1319 req->rq_phase = RQ_PHASE_RPC;
1321 spin_lock_irqsave(&imp->imp_lock, flags);
1322 req->rq_import_generation = imp->imp_generation;
1324 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1325 list_del(&req->rq_list);
1327 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1328 spin_unlock_irqrestore(&imp->imp_lock, flags);
1330 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1332 ptlrpc_import_state_name(req->rq_send_state),
1333 ptlrpc_import_state_name(imp->imp_state));
1334 lwi = LWI_INTR(interrupted_request, req);
1335 rc = l_wait_event(req->rq_reply_waitq,
1336 (req->rq_send_state == imp->imp_state ||
1339 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d == 1)",
1341 ptlrpc_import_state_name(imp->imp_state),
1342 ptlrpc_import_state_name(req->rq_send_state),
1345 spin_lock_irqsave(&imp->imp_lock, flags);
1346 list_del_init(&req->rq_list);
1351 else if (req->rq_intr) {
1354 else if (req->rq_no_resend) {
1355 spin_unlock_irqrestore(&imp->imp_lock, flags);
1356 GOTO(out, rc = -ETIMEDOUT);
1364 list_del_init(&req->rq_list);
1365 spin_unlock_irqrestore(&imp->imp_lock, flags);
1366 req->rq_status = rc; // XXX this ok?
1370 if (req->rq_resend) {
1371 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1373 if (req->rq_bulk != NULL)
1374 ptlrpc_unregister_bulk (req);
1376 DEBUG_REQ(D_HA, req, "resending: ");
1379 /* XXX this is the same as ptlrpc_set_wait */
1380 LASSERT(list_empty(&req->rq_list));
1381 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1382 spin_unlock_irqrestore(&imp->imp_lock, flags);
1384 rc = ptl_send_rpc(req);
1386 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1389 timeout = MAX(req->rq_timeout * HZ, 1);
1390 DEBUG_REQ(D_NET, req, "-- sleeping");
1392 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1394 l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1395 DEBUG_REQ(D_NET, req, "-- done sleeping");
1397 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:opc "
1398 "%s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
1399 imp->imp_obd->obd_uuid.uuid,
1400 req->rq_reqmsg->status, req->rq_xid,
1401 imp->imp_connection->c_peer.peer_ni->pni_name,
1402 imp->imp_connection->c_peer.peer_nid,
1403 req->rq_reqmsg->opc);
1405 spin_lock_irqsave(&imp->imp_lock, flags);
1406 list_del_init(&req->rq_list);
1407 spin_unlock_irqrestore(&imp->imp_lock, flags);
1409 /* If the reply was received normally, this just grabs the spinlock
1410 * (ensuring the reply callback has returned), sees that
1411 * req->rq_receiving_reply is clear and returns. */
1412 ptlrpc_unregister_reply (req);
1415 GOTO(out, rc = -EIO);
1417 /* Resend if we need to, unless we were interrupted. */
1418 if (req->rq_resend && !req->rq_intr) {
1419 /* ...unless we were specifically told otherwise. */
1420 if (req->rq_no_resend)
1421 GOTO(out, rc = -ETIMEDOUT);
1422 spin_lock_irqsave(&imp->imp_lock, flags);
1427 /* Should only be interrupted if we timed out. */
1428 if (!req->rq_timedout)
1429 DEBUG_REQ(D_ERROR, req,
1430 "rq_intr set but rq_timedout not");
1431 GOTO(out, rc = -EINTR);
1434 if (req->rq_timedout) { /* non-recoverable timeout */
1435 GOTO(out, rc = -ETIMEDOUT);
1438 if (!req->rq_replied) {
1439 /* How can this be? -eeb */
1440 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1442 GOTO(out, rc = req->rq_status);
1445 rc = after_reply (req);
1446 /* NB may return +ve success rc */
1447 if (req->rq_resend) {
1448 spin_lock_irqsave(&imp->imp_lock, flags);
1453 if (req->rq_bulk != NULL) {
1455 /* success so far. Note that anything going wrong
1456 * with bulk now, is EXTREMELY strange, since the
1457 * server must have believed that the bulk
1458 * tranferred OK before she replied with success to
1460 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1461 brc = l_wait_event(req->rq_reply_waitq,
1462 !ptlrpc_bulk_active(req->rq_bulk),
1464 LASSERT(brc == 0 || brc == -ETIMEDOUT);
1466 LASSERT(brc == -ETIMEDOUT);
1467 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1469 } else if (!req->rq_bulk->bd_success) {
1470 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1475 ptlrpc_unregister_bulk (req);
1478 LASSERT(!req->rq_receiving_reply);
1479 req->rq_phase = RQ_PHASE_INTERPRET;
1481 atomic_dec(&imp->imp_inflight);
1482 wake_up(&imp->imp_recovery_waitq);
1486 struct ptlrpc_replay_async_args {
1488 int praa_old_status;
1491 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1492 void * data, int rc)
1494 struct ptlrpc_replay_async_args *aa = data;
1495 struct obd_import *imp = req->rq_import;
1496 unsigned long flags;
1498 atomic_dec(&imp->imp_replay_inflight);
1500 if (!req->rq_replied) {
1501 CERROR("request replay timed out, restarting recovery\n");
1502 GOTO(out, rc = -ETIMEDOUT);
1506 /* Clear reply swab mask; this is a new reply in sender's byte order */
1507 req->rq_rep_swab_mask = 0;
1509 LASSERT (req->rq_nob_received <= req->rq_replen);
1510 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
1512 CERROR("unpack_rep failed: %d\n", rc);
1513 GOTO(out, rc = -EPROTO);
1516 if (req->rq_repmsg->type == PTL_RPC_MSG_ERR &&
1517 req->rq_repmsg->status == -ENOTCONN)
1518 GOTO(out, rc = req->rq_repmsg->status);
1520 /* The transno had better not change over replay. */
1521 LASSERT(req->rq_reqmsg->transno == req->rq_repmsg->transno);
1523 DEBUG_REQ(D_HA, req, "got rep");
1525 /* let the callback do fixups, possibly including in the request */
1526 if (req->rq_replay_cb)
1527 req->rq_replay_cb(req);
1529 if (req->rq_replied && req->rq_repmsg->status != aa->praa_old_status) {
1530 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1531 req->rq_repmsg->status, aa->praa_old_status);
1533 /* Put it back for re-replay. */
1534 req->rq_repmsg->status = aa->praa_old_status;
1537 spin_lock_irqsave(&imp->imp_lock, flags);
1538 imp->imp_last_replay_transno = req->rq_transno;
1539 spin_unlock_irqrestore(&imp->imp_lock, flags);
1541 /* continue with recovery */
1542 rc = ptlrpc_import_recovery_state_machine(imp);
1544 req->rq_send_state = aa->praa_old_state;
1547 /* this replay failed, so restart recovery */
1548 ptlrpc_connect_import(imp, NULL);
1554 int ptlrpc_replay_req(struct ptlrpc_request *req)
1556 struct ptlrpc_replay_async_args *aa;
1559 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1561 /* Not handling automatic bulk replay yet (or ever?) */
1562 LASSERT(req->rq_bulk == NULL);
1564 DEBUG_REQ(D_HA, req, "REPLAY");
1566 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1567 aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1568 memset(aa, 0, sizeof *aa);
1570 /* Prepare request to be resent with ptlrpcd */
1571 aa->praa_old_state = req->rq_send_state;
1572 req->rq_send_state = LUSTRE_IMP_REPLAY;
1573 req->rq_phase = RQ_PHASE_NEW;
1575 * Q: "How can a req get on the replay list if it wasn't replied?"
1576 * A: "If we failed during the replay of this request, it will still
1577 * be on the list, but rq_replied will have been reset to 0."
1579 if (req->rq_replied) {
1580 aa->praa_old_status = req->rq_repmsg->status;
1582 req->rq_replied = 0;
1585 req->rq_interpret_reply = ptlrpc_replay_interpret;
1586 atomic_inc(&req->rq_import->imp_replay_inflight);
1587 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1589 ptlrpcd_add_req(req);
1593 void ptlrpc_abort_inflight(struct obd_import *imp)
1595 unsigned long flags;
1596 struct list_head *tmp, *n;
1599 /* Make sure that no new requests get processed for this import.
1600 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1601 * this flag and then putting requests on sending_list or delayed_list.
1603 spin_lock_irqsave(&imp->imp_lock, flags);
1605 /* XXX locking? Maybe we should remove each request with the list
1606 * locked? Also, how do we know if the requests on the list are
1607 * being freed at this time?
1609 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1610 struct ptlrpc_request *req =
1611 list_entry(tmp, struct ptlrpc_request, rq_list);
1613 DEBUG_REQ(D_HA, req, "inflight");
1615 spin_lock (&req->rq_lock);
1616 if (req->rq_import_generation < imp->imp_generation) {
1618 ptlrpc_wake_client_req(req);
1620 spin_unlock (&req->rq_lock);
1623 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1624 struct ptlrpc_request *req =
1625 list_entry(tmp, struct ptlrpc_request, rq_list);
1627 DEBUG_REQ(D_HA, req, "aborting waiting req");
1629 spin_lock (&req->rq_lock);
1630 if (req->rq_import_generation < imp->imp_generation) {
1632 ptlrpc_wake_client_req(req);
1634 spin_unlock (&req->rq_lock);
1637 /* Last chance to free reqs left on the replay list, but we
1638 * will still leak reqs that haven't comitted. */
1639 if (imp->imp_replayable)
1640 ptlrpc_free_committed(imp);
1642 spin_unlock_irqrestore(&imp->imp_lock, flags);
1647 static __u64 ptlrpc_last_xid = 0;
1648 static spinlock_t ptlrpc_last_xid_lock = SPIN_LOCK_UNLOCKED;
1650 __u64 ptlrpc_next_xid(void)
1653 spin_lock(&ptlrpc_last_xid_lock);
1654 tmp = ++ptlrpc_last_xid;
1655 spin_unlock(&ptlrpc_last_xid_lock);